Loading...
1/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 */
6#include <linux/rcupdate.h>
7#include <linux/rculist.h>
8#include <linux/kernel.h>
9#include <linux/export.h>
10#include <linux/percpu.h>
11#include <linux/init.h>
12#include <linux/gfp.h>
13#include <linux/smp.h>
14#include <linux/cpu.h>
15
16#include "smpboot.h"
17
18enum {
19 CSD_FLAG_LOCK = 0x01,
20 CSD_FLAG_WAIT = 0x02,
21};
22
23struct call_function_data {
24 struct call_single_data __percpu *csd;
25 cpumask_var_t cpumask;
26};
27
28static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
29
30static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
31
32static int
33hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
34{
35 long cpu = (long)hcpu;
36 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
37
38 switch (action) {
39 case CPU_UP_PREPARE:
40 case CPU_UP_PREPARE_FROZEN:
41 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
42 cpu_to_node(cpu)))
43 return notifier_from_errno(-ENOMEM);
44 cfd->csd = alloc_percpu(struct call_single_data);
45 if (!cfd->csd) {
46 free_cpumask_var(cfd->cpumask);
47 return notifier_from_errno(-ENOMEM);
48 }
49 break;
50
51#ifdef CONFIG_HOTPLUG_CPU
52 case CPU_UP_CANCELED:
53 case CPU_UP_CANCELED_FROZEN:
54
55 case CPU_DEAD:
56 case CPU_DEAD_FROZEN:
57 free_cpumask_var(cfd->cpumask);
58 free_percpu(cfd->csd);
59 break;
60#endif
61 };
62
63 return NOTIFY_OK;
64}
65
66static struct notifier_block hotplug_cfd_notifier = {
67 .notifier_call = hotplug_cfd,
68};
69
70void __init call_function_init(void)
71{
72 void *cpu = (void *)(long)smp_processor_id();
73 int i;
74
75 for_each_possible_cpu(i)
76 init_llist_head(&per_cpu(call_single_queue, i));
77
78 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
79 register_cpu_notifier(&hotplug_cfd_notifier);
80}
81
82/*
83 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
84 *
85 * For non-synchronous ipi calls the csd can still be in use by the
86 * previous function call. For multi-cpu calls its even more interesting
87 * as we'll have to ensure no other cpu is observing our csd.
88 */
89static void csd_lock_wait(struct call_single_data *csd)
90{
91 while (csd->flags & CSD_FLAG_LOCK)
92 cpu_relax();
93}
94
95static void csd_lock(struct call_single_data *csd)
96{
97 csd_lock_wait(csd);
98 csd->flags |= CSD_FLAG_LOCK;
99
100 /*
101 * prevent CPU from reordering the above assignment
102 * to ->flags with any subsequent assignments to other
103 * fields of the specified call_single_data structure:
104 */
105 smp_mb();
106}
107
108static void csd_unlock(struct call_single_data *csd)
109{
110 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
111
112 /*
113 * ensure we're all done before releasing data:
114 */
115 smp_mb();
116
117 csd->flags &= ~CSD_FLAG_LOCK;
118}
119
120static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
121
122/*
123 * Insert a previously allocated call_single_data element
124 * for execution on the given CPU. data must already have
125 * ->func, ->info, and ->flags set.
126 */
127static int generic_exec_single(int cpu, struct call_single_data *csd,
128 smp_call_func_t func, void *info, int wait)
129{
130 struct call_single_data csd_stack = { .flags = 0 };
131 unsigned long flags;
132
133
134 if (cpu == smp_processor_id()) {
135 local_irq_save(flags);
136 func(info);
137 local_irq_restore(flags);
138 return 0;
139 }
140
141
142 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu))
143 return -ENXIO;
144
145
146 if (!csd) {
147 csd = &csd_stack;
148 if (!wait)
149 csd = &__get_cpu_var(csd_data);
150 }
151
152 csd_lock(csd);
153
154 csd->func = func;
155 csd->info = info;
156
157 if (wait)
158 csd->flags |= CSD_FLAG_WAIT;
159
160 /*
161 * The list addition should be visible before sending the IPI
162 * handler locks the list to pull the entry off it because of
163 * normal cache coherency rules implied by spinlocks.
164 *
165 * If IPIs can go out of order to the cache coherency protocol
166 * in an architecture, sufficient synchronisation should be added
167 * to arch code to make it appear to obey cache coherency WRT
168 * locking and barrier primitives. Generic code isn't really
169 * equipped to do the right thing...
170 */
171 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
172 arch_send_call_function_single_ipi(cpu);
173
174 if (wait)
175 csd_lock_wait(csd);
176
177 return 0;
178}
179
180/*
181 * Invoked by arch to handle an IPI for call function single. Must be
182 * called from the arch with interrupts disabled.
183 */
184void generic_smp_call_function_single_interrupt(void)
185{
186 struct llist_node *entry;
187 struct call_single_data *csd, *csd_next;
188
189 /*
190 * Shouldn't receive this interrupt on a cpu that is not yet online.
191 */
192 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
193
194 entry = llist_del_all(&__get_cpu_var(call_single_queue));
195 entry = llist_reverse_order(entry);
196
197 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
198 csd->func(csd->info);
199 csd_unlock(csd);
200 }
201}
202
203/*
204 * smp_call_function_single - Run a function on a specific CPU
205 * @func: The function to run. This must be fast and non-blocking.
206 * @info: An arbitrary pointer to pass to the function.
207 * @wait: If true, wait until function has completed on other CPUs.
208 *
209 * Returns 0 on success, else a negative status code.
210 */
211int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
212 int wait)
213{
214 int this_cpu;
215 int err;
216
217 /*
218 * prevent preemption and reschedule on another processor,
219 * as well as CPU removal
220 */
221 this_cpu = get_cpu();
222
223 /*
224 * Can deadlock when called with interrupts disabled.
225 * We allow cpu's that are not yet online though, as no one else can
226 * send smp call function interrupt to this cpu and as such deadlocks
227 * can't happen.
228 */
229 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
230 && !oops_in_progress);
231
232 err = generic_exec_single(cpu, NULL, func, info, wait);
233
234 put_cpu();
235
236 return err;
237}
238EXPORT_SYMBOL(smp_call_function_single);
239
240/**
241 * smp_call_function_single_async(): Run an asynchronous function on a
242 * specific CPU.
243 * @cpu: The CPU to run on.
244 * @csd: Pre-allocated and setup data structure
245 *
246 * Like smp_call_function_single(), but the call is asynchonous and
247 * can thus be done from contexts with disabled interrupts.
248 *
249 * The caller passes his own pre-allocated data structure
250 * (ie: embedded in an object) and is responsible for synchronizing it
251 * such that the IPIs performed on the @csd are strictly serialized.
252 *
253 * NOTE: Be careful, there is unfortunately no current debugging facility to
254 * validate the correctness of this serialization.
255 */
256int smp_call_function_single_async(int cpu, struct call_single_data *csd)
257{
258 int err = 0;
259
260 preempt_disable();
261 err = generic_exec_single(cpu, csd, csd->func, csd->info, 0);
262 preempt_enable();
263
264 return err;
265}
266EXPORT_SYMBOL_GPL(smp_call_function_single_async);
267
268/*
269 * smp_call_function_any - Run a function on any of the given cpus
270 * @mask: The mask of cpus it can run on.
271 * @func: The function to run. This must be fast and non-blocking.
272 * @info: An arbitrary pointer to pass to the function.
273 * @wait: If true, wait until function has completed.
274 *
275 * Returns 0 on success, else a negative status code (if no cpus were online).
276 *
277 * Selection preference:
278 * 1) current cpu if in @mask
279 * 2) any cpu of current node if in @mask
280 * 3) any other online cpu in @mask
281 */
282int smp_call_function_any(const struct cpumask *mask,
283 smp_call_func_t func, void *info, int wait)
284{
285 unsigned int cpu;
286 const struct cpumask *nodemask;
287 int ret;
288
289 /* Try for same CPU (cheapest) */
290 cpu = get_cpu();
291 if (cpumask_test_cpu(cpu, mask))
292 goto call;
293
294 /* Try for same node. */
295 nodemask = cpumask_of_node(cpu_to_node(cpu));
296 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
297 cpu = cpumask_next_and(cpu, nodemask, mask)) {
298 if (cpu_online(cpu))
299 goto call;
300 }
301
302 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
303 cpu = cpumask_any_and(mask, cpu_online_mask);
304call:
305 ret = smp_call_function_single(cpu, func, info, wait);
306 put_cpu();
307 return ret;
308}
309EXPORT_SYMBOL_GPL(smp_call_function_any);
310
311/**
312 * smp_call_function_many(): Run a function on a set of other CPUs.
313 * @mask: The set of cpus to run on (only runs on online subset).
314 * @func: The function to run. This must be fast and non-blocking.
315 * @info: An arbitrary pointer to pass to the function.
316 * @wait: If true, wait (atomically) until function has completed
317 * on other CPUs.
318 *
319 * If @wait is true, then returns once @func has returned.
320 *
321 * You must not call this function with disabled interrupts or from a
322 * hardware interrupt handler or from a bottom half handler. Preemption
323 * must be disabled when calling this function.
324 */
325void smp_call_function_many(const struct cpumask *mask,
326 smp_call_func_t func, void *info, bool wait)
327{
328 struct call_function_data *cfd;
329 int cpu, next_cpu, this_cpu = smp_processor_id();
330
331 /*
332 * Can deadlock when called with interrupts disabled.
333 * We allow cpu's that are not yet online though, as no one else can
334 * send smp call function interrupt to this cpu and as such deadlocks
335 * can't happen.
336 */
337 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
338 && !oops_in_progress && !early_boot_irqs_disabled);
339
340 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
341 cpu = cpumask_first_and(mask, cpu_online_mask);
342 if (cpu == this_cpu)
343 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
344
345 /* No online cpus? We're done. */
346 if (cpu >= nr_cpu_ids)
347 return;
348
349 /* Do we have another CPU which isn't us? */
350 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
351 if (next_cpu == this_cpu)
352 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
353
354 /* Fastpath: do that cpu by itself. */
355 if (next_cpu >= nr_cpu_ids) {
356 smp_call_function_single(cpu, func, info, wait);
357 return;
358 }
359
360 cfd = &__get_cpu_var(cfd_data);
361
362 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
363 cpumask_clear_cpu(this_cpu, cfd->cpumask);
364
365 /* Some callers race with other cpus changing the passed mask */
366 if (unlikely(!cpumask_weight(cfd->cpumask)))
367 return;
368
369 for_each_cpu(cpu, cfd->cpumask) {
370 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
371
372 csd_lock(csd);
373 csd->func = func;
374 csd->info = info;
375 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
376 }
377
378 /* Send a message to all CPUs in the map */
379 arch_send_call_function_ipi_mask(cfd->cpumask);
380
381 if (wait) {
382 for_each_cpu(cpu, cfd->cpumask) {
383 struct call_single_data *csd;
384
385 csd = per_cpu_ptr(cfd->csd, cpu);
386 csd_lock_wait(csd);
387 }
388 }
389}
390EXPORT_SYMBOL(smp_call_function_many);
391
392/**
393 * smp_call_function(): Run a function on all other CPUs.
394 * @func: The function to run. This must be fast and non-blocking.
395 * @info: An arbitrary pointer to pass to the function.
396 * @wait: If true, wait (atomically) until function has completed
397 * on other CPUs.
398 *
399 * Returns 0.
400 *
401 * If @wait is true, then returns once @func has returned; otherwise
402 * it returns just before the target cpu calls @func.
403 *
404 * You must not call this function with disabled interrupts or from a
405 * hardware interrupt handler or from a bottom half handler.
406 */
407int smp_call_function(smp_call_func_t func, void *info, int wait)
408{
409 preempt_disable();
410 smp_call_function_many(cpu_online_mask, func, info, wait);
411 preempt_enable();
412
413 return 0;
414}
415EXPORT_SYMBOL(smp_call_function);
416
417/* Setup configured maximum number of CPUs to activate */
418unsigned int setup_max_cpus = NR_CPUS;
419EXPORT_SYMBOL(setup_max_cpus);
420
421
422/*
423 * Setup routine for controlling SMP activation
424 *
425 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
426 * activation entirely (the MPS table probe still happens, though).
427 *
428 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
429 * greater than 0, limits the maximum number of CPUs activated in
430 * SMP mode to <NUM>.
431 */
432
433void __weak arch_disable_smp_support(void) { }
434
435static int __init nosmp(char *str)
436{
437 setup_max_cpus = 0;
438 arch_disable_smp_support();
439
440 return 0;
441}
442
443early_param("nosmp", nosmp);
444
445/* this is hard limit */
446static int __init nrcpus(char *str)
447{
448 int nr_cpus;
449
450 get_option(&str, &nr_cpus);
451 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
452 nr_cpu_ids = nr_cpus;
453
454 return 0;
455}
456
457early_param("nr_cpus", nrcpus);
458
459static int __init maxcpus(char *str)
460{
461 get_option(&str, &setup_max_cpus);
462 if (setup_max_cpus == 0)
463 arch_disable_smp_support();
464
465 return 0;
466}
467
468early_param("maxcpus", maxcpus);
469
470/* Setup number of possible processor ids */
471int nr_cpu_ids __read_mostly = NR_CPUS;
472EXPORT_SYMBOL(nr_cpu_ids);
473
474/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
475void __init setup_nr_cpu_ids(void)
476{
477 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
478}
479
480void __weak smp_announce(void)
481{
482 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
483}
484
485/* Called by boot processor to activate the rest. */
486void __init smp_init(void)
487{
488 unsigned int cpu;
489
490 idle_threads_init();
491
492 /* FIXME: This should be done in userspace --RR */
493 for_each_present_cpu(cpu) {
494 if (num_online_cpus() >= setup_max_cpus)
495 break;
496 if (!cpu_online(cpu))
497 cpu_up(cpu);
498 }
499
500 /* Any cleanup work */
501 smp_announce();
502 smp_cpus_done(setup_max_cpus);
503}
504
505/*
506 * Call a function on all processors. May be used during early boot while
507 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
508 * of local_irq_disable/enable().
509 */
510int on_each_cpu(void (*func) (void *info), void *info, int wait)
511{
512 unsigned long flags;
513 int ret = 0;
514
515 preempt_disable();
516 ret = smp_call_function(func, info, wait);
517 local_irq_save(flags);
518 func(info);
519 local_irq_restore(flags);
520 preempt_enable();
521 return ret;
522}
523EXPORT_SYMBOL(on_each_cpu);
524
525/**
526 * on_each_cpu_mask(): Run a function on processors specified by
527 * cpumask, which may include the local processor.
528 * @mask: The set of cpus to run on (only runs on online subset).
529 * @func: The function to run. This must be fast and non-blocking.
530 * @info: An arbitrary pointer to pass to the function.
531 * @wait: If true, wait (atomically) until function has completed
532 * on other CPUs.
533 *
534 * If @wait is true, then returns once @func has returned.
535 *
536 * You must not call this function with disabled interrupts or from a
537 * hardware interrupt handler or from a bottom half handler. The
538 * exception is that it may be used during early boot while
539 * early_boot_irqs_disabled is set.
540 */
541void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
542 void *info, bool wait)
543{
544 int cpu = get_cpu();
545
546 smp_call_function_many(mask, func, info, wait);
547 if (cpumask_test_cpu(cpu, mask)) {
548 unsigned long flags;
549 local_irq_save(flags);
550 func(info);
551 local_irq_restore(flags);
552 }
553 put_cpu();
554}
555EXPORT_SYMBOL(on_each_cpu_mask);
556
557/*
558 * on_each_cpu_cond(): Call a function on each processor for which
559 * the supplied function cond_func returns true, optionally waiting
560 * for all the required CPUs to finish. This may include the local
561 * processor.
562 * @cond_func: A callback function that is passed a cpu id and
563 * the the info parameter. The function is called
564 * with preemption disabled. The function should
565 * return a blooean value indicating whether to IPI
566 * the specified CPU.
567 * @func: The function to run on all applicable CPUs.
568 * This must be fast and non-blocking.
569 * @info: An arbitrary pointer to pass to both functions.
570 * @wait: If true, wait (atomically) until function has
571 * completed on other CPUs.
572 * @gfp_flags: GFP flags to use when allocating the cpumask
573 * used internally by the function.
574 *
575 * The function might sleep if the GFP flags indicates a non
576 * atomic allocation is allowed.
577 *
578 * Preemption is disabled to protect against CPUs going offline but not online.
579 * CPUs going online during the call will not be seen or sent an IPI.
580 *
581 * You must not call this function with disabled interrupts or
582 * from a hardware interrupt handler or from a bottom half handler.
583 */
584void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
585 smp_call_func_t func, void *info, bool wait,
586 gfp_t gfp_flags)
587{
588 cpumask_var_t cpus;
589 int cpu, ret;
590
591 might_sleep_if(gfp_flags & __GFP_WAIT);
592
593 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
594 preempt_disable();
595 for_each_online_cpu(cpu)
596 if (cond_func(cpu, info))
597 cpumask_set_cpu(cpu, cpus);
598 on_each_cpu_mask(cpus, func, info, wait);
599 preempt_enable();
600 free_cpumask_var(cpus);
601 } else {
602 /*
603 * No free cpumask, bother. No matter, we'll
604 * just have to IPI them one by one.
605 */
606 preempt_disable();
607 for_each_online_cpu(cpu)
608 if (cond_func(cpu, info)) {
609 ret = smp_call_function_single(cpu, func,
610 info, wait);
611 WARN_ON_ONCE(!ret);
612 }
613 preempt_enable();
614 }
615}
616EXPORT_SYMBOL(on_each_cpu_cond);
617
618static void do_nothing(void *unused)
619{
620}
621
622/**
623 * kick_all_cpus_sync - Force all cpus out of idle
624 *
625 * Used to synchronize the update of pm_idle function pointer. It's
626 * called after the pointer is updated and returns after the dummy
627 * callback function has been executed on all cpus. The execution of
628 * the function can only happen on the remote cpus after they have
629 * left the idle function which had been called via pm_idle function
630 * pointer. So it's guaranteed that nothing uses the previous pointer
631 * anymore.
632 */
633void kick_all_cpus_sync(void)
634{
635 /* Make sure the change is visible before we kick the cpus */
636 smp_mb();
637 smp_call_function(do_nothing, NULL, 1);
638}
639EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
1/*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 */
6
7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
9#include <linux/irq_work.h>
10#include <linux/rcupdate.h>
11#include <linux/rculist.h>
12#include <linux/kernel.h>
13#include <linux/export.h>
14#include <linux/percpu.h>
15#include <linux/init.h>
16#include <linux/gfp.h>
17#include <linux/smp.h>
18#include <linux/cpu.h>
19#include <linux/sched.h>
20#include <linux/sched/idle.h>
21#include <linux/hypervisor.h>
22
23#include "smpboot.h"
24
25enum {
26 CSD_FLAG_LOCK = 0x01,
27 CSD_FLAG_SYNCHRONOUS = 0x02,
28};
29
30struct call_function_data {
31 call_single_data_t __percpu *csd;
32 cpumask_var_t cpumask;
33 cpumask_var_t cpumask_ipi;
34};
35
36static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
37
38static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
39
40static void flush_smp_call_function_queue(bool warn_cpu_offline);
41
42int smpcfd_prepare_cpu(unsigned int cpu)
43{
44 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
45
46 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
47 cpu_to_node(cpu)))
48 return -ENOMEM;
49 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
50 cpu_to_node(cpu))) {
51 free_cpumask_var(cfd->cpumask);
52 return -ENOMEM;
53 }
54 cfd->csd = alloc_percpu(call_single_data_t);
55 if (!cfd->csd) {
56 free_cpumask_var(cfd->cpumask);
57 free_cpumask_var(cfd->cpumask_ipi);
58 return -ENOMEM;
59 }
60
61 return 0;
62}
63
64int smpcfd_dead_cpu(unsigned int cpu)
65{
66 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
67
68 free_cpumask_var(cfd->cpumask);
69 free_cpumask_var(cfd->cpumask_ipi);
70 free_percpu(cfd->csd);
71 return 0;
72}
73
74int smpcfd_dying_cpu(unsigned int cpu)
75{
76 /*
77 * The IPIs for the smp-call-function callbacks queued by other
78 * CPUs might arrive late, either due to hardware latencies or
79 * because this CPU disabled interrupts (inside stop-machine)
80 * before the IPIs were sent. So flush out any pending callbacks
81 * explicitly (without waiting for the IPIs to arrive), to
82 * ensure that the outgoing CPU doesn't go offline with work
83 * still pending.
84 */
85 flush_smp_call_function_queue(false);
86 return 0;
87}
88
89void __init call_function_init(void)
90{
91 int i;
92
93 for_each_possible_cpu(i)
94 init_llist_head(&per_cpu(call_single_queue, i));
95
96 smpcfd_prepare_cpu(smp_processor_id());
97}
98
99/*
100 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
101 *
102 * For non-synchronous ipi calls the csd can still be in use by the
103 * previous function call. For multi-cpu calls its even more interesting
104 * as we'll have to ensure no other cpu is observing our csd.
105 */
106static __always_inline void csd_lock_wait(call_single_data_t *csd)
107{
108 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
109}
110
111static __always_inline void csd_lock(call_single_data_t *csd)
112{
113 csd_lock_wait(csd);
114 csd->flags |= CSD_FLAG_LOCK;
115
116 /*
117 * prevent CPU from reordering the above assignment
118 * to ->flags with any subsequent assignments to other
119 * fields of the specified call_single_data_t structure:
120 */
121 smp_wmb();
122}
123
124static __always_inline void csd_unlock(call_single_data_t *csd)
125{
126 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
127
128 /*
129 * ensure we're all done before releasing data:
130 */
131 smp_store_release(&csd->flags, 0);
132}
133
134static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
135
136/*
137 * Insert a previously allocated call_single_data_t element
138 * for execution on the given CPU. data must already have
139 * ->func, ->info, and ->flags set.
140 */
141static int generic_exec_single(int cpu, call_single_data_t *csd,
142 smp_call_func_t func, void *info)
143{
144 if (cpu == smp_processor_id()) {
145 unsigned long flags;
146
147 /*
148 * We can unlock early even for the synchronous on-stack case,
149 * since we're doing this from the same CPU..
150 */
151 csd_unlock(csd);
152 local_irq_save(flags);
153 func(info);
154 local_irq_restore(flags);
155 return 0;
156 }
157
158
159 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
160 csd_unlock(csd);
161 return -ENXIO;
162 }
163
164 csd->func = func;
165 csd->info = info;
166
167 /*
168 * The list addition should be visible before sending the IPI
169 * handler locks the list to pull the entry off it because of
170 * normal cache coherency rules implied by spinlocks.
171 *
172 * If IPIs can go out of order to the cache coherency protocol
173 * in an architecture, sufficient synchronisation should be added
174 * to arch code to make it appear to obey cache coherency WRT
175 * locking and barrier primitives. Generic code isn't really
176 * equipped to do the right thing...
177 */
178 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
179 arch_send_call_function_single_ipi(cpu);
180
181 return 0;
182}
183
184/**
185 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
186 *
187 * Invoked by arch to handle an IPI for call function single.
188 * Must be called with interrupts disabled.
189 */
190void generic_smp_call_function_single_interrupt(void)
191{
192 flush_smp_call_function_queue(true);
193}
194
195/**
196 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
197 *
198 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
199 * offline CPU. Skip this check if set to 'false'.
200 *
201 * Flush any pending smp-call-function callbacks queued on this CPU. This is
202 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
203 * to ensure that all pending IPI callbacks are run before it goes completely
204 * offline.
205 *
206 * Loop through the call_single_queue and run all the queued callbacks.
207 * Must be called with interrupts disabled.
208 */
209static void flush_smp_call_function_queue(bool warn_cpu_offline)
210{
211 struct llist_head *head;
212 struct llist_node *entry;
213 call_single_data_t *csd, *csd_next;
214 static bool warned;
215
216 lockdep_assert_irqs_disabled();
217
218 head = this_cpu_ptr(&call_single_queue);
219 entry = llist_del_all(head);
220 entry = llist_reverse_order(entry);
221
222 /* There shouldn't be any pending callbacks on an offline CPU. */
223 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
224 !warned && !llist_empty(head))) {
225 warned = true;
226 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
227
228 /*
229 * We don't have to use the _safe() variant here
230 * because we are not invoking the IPI handlers yet.
231 */
232 llist_for_each_entry(csd, entry, llist)
233 pr_warn("IPI callback %pS sent to offline CPU\n",
234 csd->func);
235 }
236
237 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
238 smp_call_func_t func = csd->func;
239 void *info = csd->info;
240
241 /* Do we wait until *after* callback? */
242 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
243 func(info);
244 csd_unlock(csd);
245 } else {
246 csd_unlock(csd);
247 func(info);
248 }
249 }
250
251 /*
252 * Handle irq works queued remotely by irq_work_queue_on().
253 * Smp functions above are typically synchronous so they
254 * better run first since some other CPUs may be busy waiting
255 * for them.
256 */
257 irq_work_run();
258}
259
260/*
261 * smp_call_function_single - Run a function on a specific CPU
262 * @func: The function to run. This must be fast and non-blocking.
263 * @info: An arbitrary pointer to pass to the function.
264 * @wait: If true, wait until function has completed on other CPUs.
265 *
266 * Returns 0 on success, else a negative status code.
267 */
268int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
269 int wait)
270{
271 call_single_data_t *csd;
272 call_single_data_t csd_stack = {
273 .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS,
274 };
275 int this_cpu;
276 int err;
277
278 /*
279 * prevent preemption and reschedule on another processor,
280 * as well as CPU removal
281 */
282 this_cpu = get_cpu();
283
284 /*
285 * Can deadlock when called with interrupts disabled.
286 * We allow cpu's that are not yet online though, as no one else can
287 * send smp call function interrupt to this cpu and as such deadlocks
288 * can't happen.
289 */
290 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
291 && !oops_in_progress);
292
293 csd = &csd_stack;
294 if (!wait) {
295 csd = this_cpu_ptr(&csd_data);
296 csd_lock(csd);
297 }
298
299 err = generic_exec_single(cpu, csd, func, info);
300
301 if (wait)
302 csd_lock_wait(csd);
303
304 put_cpu();
305
306 return err;
307}
308EXPORT_SYMBOL(smp_call_function_single);
309
310/**
311 * smp_call_function_single_async(): Run an asynchronous function on a
312 * specific CPU.
313 * @cpu: The CPU to run on.
314 * @csd: Pre-allocated and setup data structure
315 *
316 * Like smp_call_function_single(), but the call is asynchonous and
317 * can thus be done from contexts with disabled interrupts.
318 *
319 * The caller passes his own pre-allocated data structure
320 * (ie: embedded in an object) and is responsible for synchronizing it
321 * such that the IPIs performed on the @csd are strictly serialized.
322 *
323 * NOTE: Be careful, there is unfortunately no current debugging facility to
324 * validate the correctness of this serialization.
325 */
326int smp_call_function_single_async(int cpu, call_single_data_t *csd)
327{
328 int err = 0;
329
330 preempt_disable();
331
332 /* We could deadlock if we have to wait here with interrupts disabled! */
333 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
334 csd_lock_wait(csd);
335
336 csd->flags = CSD_FLAG_LOCK;
337 smp_wmb();
338
339 err = generic_exec_single(cpu, csd, csd->func, csd->info);
340 preempt_enable();
341
342 return err;
343}
344EXPORT_SYMBOL_GPL(smp_call_function_single_async);
345
346/*
347 * smp_call_function_any - Run a function on any of the given cpus
348 * @mask: The mask of cpus it can run on.
349 * @func: The function to run. This must be fast and non-blocking.
350 * @info: An arbitrary pointer to pass to the function.
351 * @wait: If true, wait until function has completed.
352 *
353 * Returns 0 on success, else a negative status code (if no cpus were online).
354 *
355 * Selection preference:
356 * 1) current cpu if in @mask
357 * 2) any cpu of current node if in @mask
358 * 3) any other online cpu in @mask
359 */
360int smp_call_function_any(const struct cpumask *mask,
361 smp_call_func_t func, void *info, int wait)
362{
363 unsigned int cpu;
364 const struct cpumask *nodemask;
365 int ret;
366
367 /* Try for same CPU (cheapest) */
368 cpu = get_cpu();
369 if (cpumask_test_cpu(cpu, mask))
370 goto call;
371
372 /* Try for same node. */
373 nodemask = cpumask_of_node(cpu_to_node(cpu));
374 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
375 cpu = cpumask_next_and(cpu, nodemask, mask)) {
376 if (cpu_online(cpu))
377 goto call;
378 }
379
380 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
381 cpu = cpumask_any_and(mask, cpu_online_mask);
382call:
383 ret = smp_call_function_single(cpu, func, info, wait);
384 put_cpu();
385 return ret;
386}
387EXPORT_SYMBOL_GPL(smp_call_function_any);
388
389/**
390 * smp_call_function_many(): Run a function on a set of other CPUs.
391 * @mask: The set of cpus to run on (only runs on online subset).
392 * @func: The function to run. This must be fast and non-blocking.
393 * @info: An arbitrary pointer to pass to the function.
394 * @wait: If true, wait (atomically) until function has completed
395 * on other CPUs.
396 *
397 * If @wait is true, then returns once @func has returned.
398 *
399 * You must not call this function with disabled interrupts or from a
400 * hardware interrupt handler or from a bottom half handler. Preemption
401 * must be disabled when calling this function.
402 */
403void smp_call_function_many(const struct cpumask *mask,
404 smp_call_func_t func, void *info, bool wait)
405{
406 struct call_function_data *cfd;
407 int cpu, next_cpu, this_cpu = smp_processor_id();
408
409 /*
410 * Can deadlock when called with interrupts disabled.
411 * We allow cpu's that are not yet online though, as no one else can
412 * send smp call function interrupt to this cpu and as such deadlocks
413 * can't happen.
414 */
415 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
416 && !oops_in_progress && !early_boot_irqs_disabled);
417
418 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
419 cpu = cpumask_first_and(mask, cpu_online_mask);
420 if (cpu == this_cpu)
421 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
422
423 /* No online cpus? We're done. */
424 if (cpu >= nr_cpu_ids)
425 return;
426
427 /* Do we have another CPU which isn't us? */
428 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
429 if (next_cpu == this_cpu)
430 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
431
432 /* Fastpath: do that cpu by itself. */
433 if (next_cpu >= nr_cpu_ids) {
434 smp_call_function_single(cpu, func, info, wait);
435 return;
436 }
437
438 cfd = this_cpu_ptr(&cfd_data);
439
440 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
441 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
442
443 /* Some callers race with other cpus changing the passed mask */
444 if (unlikely(!cpumask_weight(cfd->cpumask)))
445 return;
446
447 cpumask_clear(cfd->cpumask_ipi);
448 for_each_cpu(cpu, cfd->cpumask) {
449 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
450
451 csd_lock(csd);
452 if (wait)
453 csd->flags |= CSD_FLAG_SYNCHRONOUS;
454 csd->func = func;
455 csd->info = info;
456 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
457 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
458 }
459
460 /* Send a message to all CPUs in the map */
461 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
462
463 if (wait) {
464 for_each_cpu(cpu, cfd->cpumask) {
465 call_single_data_t *csd;
466
467 csd = per_cpu_ptr(cfd->csd, cpu);
468 csd_lock_wait(csd);
469 }
470 }
471}
472EXPORT_SYMBOL(smp_call_function_many);
473
474/**
475 * smp_call_function(): Run a function on all other CPUs.
476 * @func: The function to run. This must be fast and non-blocking.
477 * @info: An arbitrary pointer to pass to the function.
478 * @wait: If true, wait (atomically) until function has completed
479 * on other CPUs.
480 *
481 * Returns 0.
482 *
483 * If @wait is true, then returns once @func has returned; otherwise
484 * it returns just before the target cpu calls @func.
485 *
486 * You must not call this function with disabled interrupts or from a
487 * hardware interrupt handler or from a bottom half handler.
488 */
489int smp_call_function(smp_call_func_t func, void *info, int wait)
490{
491 preempt_disable();
492 smp_call_function_many(cpu_online_mask, func, info, wait);
493 preempt_enable();
494
495 return 0;
496}
497EXPORT_SYMBOL(smp_call_function);
498
499/* Setup configured maximum number of CPUs to activate */
500unsigned int setup_max_cpus = NR_CPUS;
501EXPORT_SYMBOL(setup_max_cpus);
502
503
504/*
505 * Setup routine for controlling SMP activation
506 *
507 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
508 * activation entirely (the MPS table probe still happens, though).
509 *
510 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
511 * greater than 0, limits the maximum number of CPUs activated in
512 * SMP mode to <NUM>.
513 */
514
515void __weak arch_disable_smp_support(void) { }
516
517static int __init nosmp(char *str)
518{
519 setup_max_cpus = 0;
520 arch_disable_smp_support();
521
522 return 0;
523}
524
525early_param("nosmp", nosmp);
526
527/* this is hard limit */
528static int __init nrcpus(char *str)
529{
530 int nr_cpus;
531
532 get_option(&str, &nr_cpus);
533 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
534 nr_cpu_ids = nr_cpus;
535
536 return 0;
537}
538
539early_param("nr_cpus", nrcpus);
540
541static int __init maxcpus(char *str)
542{
543 get_option(&str, &setup_max_cpus);
544 if (setup_max_cpus == 0)
545 arch_disable_smp_support();
546
547 return 0;
548}
549
550early_param("maxcpus", maxcpus);
551
552/* Setup number of possible processor ids */
553unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
554EXPORT_SYMBOL(nr_cpu_ids);
555
556/* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
557void __init setup_nr_cpu_ids(void)
558{
559 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
560}
561
562/* Called by boot processor to activate the rest. */
563void __init smp_init(void)
564{
565 int num_nodes, num_cpus;
566 unsigned int cpu;
567
568 idle_threads_init();
569 cpuhp_threads_init();
570
571 pr_info("Bringing up secondary CPUs ...\n");
572
573 /* FIXME: This should be done in userspace --RR */
574 for_each_present_cpu(cpu) {
575 if (num_online_cpus() >= setup_max_cpus)
576 break;
577 if (!cpu_online(cpu))
578 cpu_up(cpu);
579 }
580
581 num_nodes = num_online_nodes();
582 num_cpus = num_online_cpus();
583 pr_info("Brought up %d node%s, %d CPU%s\n",
584 num_nodes, (num_nodes > 1 ? "s" : ""),
585 num_cpus, (num_cpus > 1 ? "s" : ""));
586
587 /* Any cleanup work */
588 smp_cpus_done(setup_max_cpus);
589}
590
591/*
592 * Call a function on all processors. May be used during early boot while
593 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
594 * of local_irq_disable/enable().
595 */
596int on_each_cpu(void (*func) (void *info), void *info, int wait)
597{
598 unsigned long flags;
599 int ret = 0;
600
601 preempt_disable();
602 ret = smp_call_function(func, info, wait);
603 local_irq_save(flags);
604 func(info);
605 local_irq_restore(flags);
606 preempt_enable();
607 return ret;
608}
609EXPORT_SYMBOL(on_each_cpu);
610
611/**
612 * on_each_cpu_mask(): Run a function on processors specified by
613 * cpumask, which may include the local processor.
614 * @mask: The set of cpus to run on (only runs on online subset).
615 * @func: The function to run. This must be fast and non-blocking.
616 * @info: An arbitrary pointer to pass to the function.
617 * @wait: If true, wait (atomically) until function has completed
618 * on other CPUs.
619 *
620 * If @wait is true, then returns once @func has returned.
621 *
622 * You must not call this function with disabled interrupts or from a
623 * hardware interrupt handler or from a bottom half handler. The
624 * exception is that it may be used during early boot while
625 * early_boot_irqs_disabled is set.
626 */
627void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
628 void *info, bool wait)
629{
630 int cpu = get_cpu();
631
632 smp_call_function_many(mask, func, info, wait);
633 if (cpumask_test_cpu(cpu, mask)) {
634 unsigned long flags;
635 local_irq_save(flags);
636 func(info);
637 local_irq_restore(flags);
638 }
639 put_cpu();
640}
641EXPORT_SYMBOL(on_each_cpu_mask);
642
643/*
644 * on_each_cpu_cond(): Call a function on each processor for which
645 * the supplied function cond_func returns true, optionally waiting
646 * for all the required CPUs to finish. This may include the local
647 * processor.
648 * @cond_func: A callback function that is passed a cpu id and
649 * the the info parameter. The function is called
650 * with preemption disabled. The function should
651 * return a blooean value indicating whether to IPI
652 * the specified CPU.
653 * @func: The function to run on all applicable CPUs.
654 * This must be fast and non-blocking.
655 * @info: An arbitrary pointer to pass to both functions.
656 * @wait: If true, wait (atomically) until function has
657 * completed on other CPUs.
658 * @gfp_flags: GFP flags to use when allocating the cpumask
659 * used internally by the function.
660 *
661 * The function might sleep if the GFP flags indicates a non
662 * atomic allocation is allowed.
663 *
664 * Preemption is disabled to protect against CPUs going offline but not online.
665 * CPUs going online during the call will not be seen or sent an IPI.
666 *
667 * You must not call this function with disabled interrupts or
668 * from a hardware interrupt handler or from a bottom half handler.
669 */
670void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
671 smp_call_func_t func, void *info, bool wait,
672 gfp_t gfp_flags)
673{
674 cpumask_var_t cpus;
675 int cpu, ret;
676
677 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
678
679 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
680 preempt_disable();
681 for_each_online_cpu(cpu)
682 if (cond_func(cpu, info))
683 cpumask_set_cpu(cpu, cpus);
684 on_each_cpu_mask(cpus, func, info, wait);
685 preempt_enable();
686 free_cpumask_var(cpus);
687 } else {
688 /*
689 * No free cpumask, bother. No matter, we'll
690 * just have to IPI them one by one.
691 */
692 preempt_disable();
693 for_each_online_cpu(cpu)
694 if (cond_func(cpu, info)) {
695 ret = smp_call_function_single(cpu, func,
696 info, wait);
697 WARN_ON_ONCE(ret);
698 }
699 preempt_enable();
700 }
701}
702EXPORT_SYMBOL(on_each_cpu_cond);
703
704static void do_nothing(void *unused)
705{
706}
707
708/**
709 * kick_all_cpus_sync - Force all cpus out of idle
710 *
711 * Used to synchronize the update of pm_idle function pointer. It's
712 * called after the pointer is updated and returns after the dummy
713 * callback function has been executed on all cpus. The execution of
714 * the function can only happen on the remote cpus after they have
715 * left the idle function which had been called via pm_idle function
716 * pointer. So it's guaranteed that nothing uses the previous pointer
717 * anymore.
718 */
719void kick_all_cpus_sync(void)
720{
721 /* Make sure the change is visible before we kick the cpus */
722 smp_mb();
723 smp_call_function(do_nothing, NULL, 1);
724}
725EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
726
727/**
728 * wake_up_all_idle_cpus - break all cpus out of idle
729 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
730 * including idle polling cpus, for non-idle cpus, we will do nothing
731 * for them.
732 */
733void wake_up_all_idle_cpus(void)
734{
735 int cpu;
736
737 preempt_disable();
738 for_each_online_cpu(cpu) {
739 if (cpu == smp_processor_id())
740 continue;
741
742 wake_up_if_idle(cpu);
743 }
744 preempt_enable();
745}
746EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
747
748/**
749 * smp_call_on_cpu - Call a function on a specific cpu
750 *
751 * Used to call a function on a specific cpu and wait for it to return.
752 * Optionally make sure the call is done on a specified physical cpu via vcpu
753 * pinning in order to support virtualized environments.
754 */
755struct smp_call_on_cpu_struct {
756 struct work_struct work;
757 struct completion done;
758 int (*func)(void *);
759 void *data;
760 int ret;
761 int cpu;
762};
763
764static void smp_call_on_cpu_callback(struct work_struct *work)
765{
766 struct smp_call_on_cpu_struct *sscs;
767
768 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
769 if (sscs->cpu >= 0)
770 hypervisor_pin_vcpu(sscs->cpu);
771 sscs->ret = sscs->func(sscs->data);
772 if (sscs->cpu >= 0)
773 hypervisor_pin_vcpu(-1);
774
775 complete(&sscs->done);
776}
777
778int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
779{
780 struct smp_call_on_cpu_struct sscs = {
781 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
782 .func = func,
783 .data = par,
784 .cpu = phys ? cpu : -1,
785 };
786
787 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
788
789 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
790 return -ENXIO;
791
792 queue_work_on(cpu, system_wq, &sscs.work);
793 wait_for_completion(&sscs.done);
794
795 return sscs.ret;
796}
797EXPORT_SYMBOL_GPL(smp_call_on_cpu);